Language Detection for Music Information Retrieval SCL Vikramjit Mitra, Daniel Garcia-Romero

advertisement
Language Detection for
Music Information Retrieval
Vikramjit Mitra, Daniel Garcia-Romero and Carol Y. Espy-Wilson
Speech Communication Lab
Introduction
This research aims to introduce a systematic
approach towards audio information retrieval using a
Vocal-Instrumental detector (VID) and
Language
Identification (LID).
Cascade architecture for language dependent Genre detection
Genre detector
Vocals
Instrumental
Detector
Audio
Vocals
The mean super-vector
Background
UBM
An Audio
File
Feature
Extraction
.
.
.
Instrumental
Genres
• Since Genre types vary with language, knowledge
of the language used in the vocals can improve
Genre detection accuracy.
• Detecting language provides content related
information that allows us to work with multilingual databases
Example of a possible Information Retrieval based Audio-browser
– Used Gaussian Mixture Model as the backend
– Two different feature sets considered
– Mel-frequency Cepstral Coefficients (MFCC)
– Shifted Delta Coefficients [1]
– UBM was trained using the entire training set
 SV-MSDC: MSDC with 256 mix. GMM (Error 8.65%)
 Error can be further ↓ by fusing super-vector LID with
GMM LID (Error 8.39%)
The Fusion model
Super-vector
SVM LID
LR Score
Normalization
Σ
SVM
 SV-MFCC: MFCC with 256 mix. GMM (Error 10.4%)
Background
UBM
Feature
Extraction
GMM
Model
2 GMM-super-vectors considered
The GMM-UBM based VID system [2]
Audio
File
µ-Adapt
λ(µ,Σ,w)
Vocals – Instrumental Detector (VID)
Chinese
Genres
Instrumental
 Additionally, we made sure that the vocals are not
dominated by nonsense words or humming.
Super-vector SVM – LID System
 67% used for training, rest for testing
English
Genres
Language
Identification
 Since we are unable to separate the vocals from the
background music, we segmented the audio
segments that contain vocals in such a way that the
vocals are dominant (i.e., the duration of the vocals is
greater than 50%)
SCL
An Audio
File
α
Feature
Extraction
Σ
Fused
Score
1-α
GMM-UBM
LID
GMM
Model
Confusion matrix for the Fused LID system
Confusion matrix for 512 order GMM-VID
Database
• Since no standard multi-language audio corpora is
available, our corpus was created in-house and
manually tagged
• 1358 audio segments of approximate duration 10
secs were created from larger audio files and
downsampled to 8 kHz (region where speech is
dominant)
Distribution of files by Language
Vocals
Instrumental
Vocals
100
0
Instrumental
12.9
87.1
The GMM Language Identifier (LID)
– Different feature sets considered:
– FSDC provides multi-resolution, but dimension ↑
Avg. Error for GMM-LID using 2048 mixture GMM-UBM
Error (%)
English Bengali Hindi Spanish Chinese Russian
#Segments
465
234
210
103
85
75
186
#Genre
7
3
3
2
2
2
2
RUS
BEN
HIN
ESP
ENG
80.00
0.00
4.00
0.00
4.00
12.00
RUS
0.00
94.12
0.00
2.94
0.00
2.94
BEN
0.00
0.00
89.74
1.28
0.00
8.97
HIN
0.00
0.00
4.29
87.14
0.00
8.57
ESP
0.00
0.00
0.00
3.57
75.00
21.43
ENG
0.00
0.00
1.29
0.00
0.00
98.71
Conclusion
– MFCC, SDC, MSDC, FSDC
Instrumental
Vocals
CHN
CHN
MFCC
SDC
MSDC
FSDC
17.3
20.7
18.3
20.2
 Vocals can be detected from Instrumentals with high acc.
 Language can be detected with an average accuracy 91.6%.
References
[1] B. Bielefeld, “Language identification using shifted delta cepstrum”, Proc.
14th Annual Speech Research Symposium, 1994
[2] D.A. Reynolds, “Comparison of background normalization methods for
text-independent speaker verification”. Proc. EU Conf. on Speech Comm. &
Tech., pp.963–966, Sep 1997.
Download